Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-Shot Cross-Lingual Document-Level Event Causality Identification with Heterogeneous Graph Contrastive Transfer Learning

Zhitao He, Pengfei Cao, Zhuoran Jin, Yubo Chen, Kang Liu, Zhiqiang Zhang, Mengshu Sun, Jun Zhao

TL;DR

This work tackles document-level event causality identification (ECI) in zero-shot cross-lingual settings, focusing on low-resource languages. It introduces GIMC, a framework that combines a heterogeneous graph interaction network with a multi-granularity contrastive transfer learning module to align cross-lingual causal representations. Through extensive experiments on MECI, GIMC achieves substantial improvements in both monolingual and multilingual scenarios, outperforming previous state-of-the-art methods by notable margins and even surpassing GPT-3.5 in few-shot settings for multilingual ECI. The approach demonstrates that modeling long-distance event dependencies with a rich graph structure, coupled with cross-lingual pattern alignment, yields robust cross-lingual transfer for complex document-level reasoning tasks.

Abstract

Event Causality Identification (ECI) refers to the detection of causal relations between events in texts. However, most existing studies focus on sentence-level ECI with high-resource languages, leaving more challenging document-level ECI (DECI) with low-resource languages under-explored. In this paper, we propose a Heterogeneous Graph Interaction Model with Multi-granularity Contrastive Transfer Learning (GIMC) for zero-shot cross-lingual document-level ECI. Specifically, we introduce a heterogeneous graph interaction network to model the long-distance dependencies between events that are scattered over a document. Then, to improve cross-lingual transferability of causal knowledge learned from the source language, we propose a multi-granularity contrastive transfer learning module to align the causal representations across languages. Extensive experiments show our framework outperforms the previous state-of-the-art model by 9.4% and 8.2% of average F1 score on monolingual and multilingual scenarios respectively. Notably, in the multilingual scenario, our zero-shot framework even exceeds GPT-3.5 with few-shot learning by 24.3% in overall performance.

Zero-Shot Cross-Lingual Document-Level Event Causality Identification with Heterogeneous Graph Contrastive Transfer Learning

TL;DR

This work tackles document-level event causality identification (ECI) in zero-shot cross-lingual settings, focusing on low-resource languages. It introduces GIMC, a framework that combines a heterogeneous graph interaction network with a multi-granularity contrastive transfer learning module to align cross-lingual causal representations. Through extensive experiments on MECI, GIMC achieves substantial improvements in both monolingual and multilingual scenarios, outperforming previous state-of-the-art methods by notable margins and even surpassing GPT-3.5 in few-shot settings for multilingual ECI. The approach demonstrates that modeling long-distance event dependencies with a rich graph structure, coupled with cross-lingual pattern alignment, yields robust cross-lingual transfer for complex document-level reasoning tasks.

Abstract

Event Causality Identification (ECI) refers to the detection of causal relations between events in texts. However, most existing studies focus on sentence-level ECI with high-resource languages, leaving more challenging document-level ECI (DECI) with low-resource languages under-explored. In this paper, we propose a Heterogeneous Graph Interaction Model with Multi-granularity Contrastive Transfer Learning (GIMC) for zero-shot cross-lingual document-level ECI. Specifically, we introduce a heterogeneous graph interaction network to model the long-distance dependencies between events that are scattered over a document. Then, to improve cross-lingual transferability of causal knowledge learned from the source language, we propose a multi-granularity contrastive transfer learning module to align the causal representations across languages. Extensive experiments show our framework outperforms the previous state-of-the-art model by 9.4% and 8.2% of average F1 score on monolingual and multilingual scenarios respectively. Notably, in the multilingual scenario, our zero-shot framework even exceeds GPT-3.5 with few-shot learning by 24.3% in overall performance.
Paper Structure (24 sections, 10 equations, 13 figures, 7 tables)

This paper contains 24 sections, 10 equations, 13 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Informative Phrase Extraction
  4. Heterogeneous Graph Interaction Network
  5. Multi-granularity Contrastive Learning Module
  6. Training
  7. Experiments
  8. Dataset
  9. Experimental Settings
  10. Baselines
  11. Main Results
  12. Effects of Heterogeneous Graph Interaction Network
  13. Effects of Cross-lingual Contrastive Transfer Learning Module
  14. Cross-lingual Transfer Effect of Different Languages
  15. Related Work
  16. ...and 9 more sections

Figures (13)

  • Figure 1: Zero-shot cross-lingual document-level event causality identification. Data is non-parallel.
  • Figure 2: We show event-relevant sentences. Events are colored and we underline the informative phrases related to colored events. Red lines and gray lines indicate the causal relations and arguments of events, respectively.
  • Figure 3: Example of informative phrase extraction
  • Figure 4: Overview of our proposed heterogeneous graph interaction network with multi-granularity contrastive transfer learning (GIMC) for zero-shot cross-lingual document-level ECI.
  • Figure 5: Few-shot prompt for English. We set language specific prompts for different source languages. The passage and relations are truncated.
  • ...and 8 more figures